What is federated learning and how does it relate to offline AI?

Federated learning trains AI models across decentralized devices without sharing raw data, making it inherently offline-friendly. Each device trains on local data and shares only model gradients or weight updates. The process: a central server distributes the base model, devices train locally for several epochs, encrypted gradient updates are aggregated (FedAvg, FedProx, or FedMA algorithms) when connectivity allows. For offline scenarios: devices accumulate training updates over days/weeks, then synchronize during periodic connectivity windows. Key challenges include non-IID data distribution across devices (solved with gradient normalization), communication efficiency (solved with gradient compression and top-k sparsification), and privacy (solved with differential privacy noise addition and secure aggregation). Google uses federated learning for Gboard next-word prediction, training on billions of devices while keeping user typing data private.